Risk-Aware Rapid Data Evacuation for Large- Scale Disasters in Optical Cloud Networks

Transcription

1 Risk-Aware Rapid Data Evacuation for Large- Scale Disasters in Optical Cloud Networks Presenter: Yongcheng (Jeremy) Li PhD student, School of Electronic and Information Engineering, Soochow University, China Group Meeting, Friday, August 26, 2016

2 Outline 1. Background 2. Risk-Aware Rapid Data Evacuation For Large- Scale Disasters 3. Heuristic Algorithm 4. Performance Evaluation 5. Conclusion 6. Future Work 2

3 Background Enterprises deploy their cloud services such as cloud data storage and applications in distributed datacenter (DC) networks. Cloud services require Terabytes or Petabytes of data transfer. Optical networks can be used to facilitate data transfers. Advantage: high bandwidth and low latency in inter-dc networks. Disadvantage: services can be disrupted by disasters (such as earthquakes, tornadoes, and intentional attacks). A large-scale disaster can lead to high data loss and service disruptions Japan Earthquake damaged many cloud providers data. 3

4 Background (contd.) Provide redundancy (and protection) against data loss. Distributed content/service replicas in different DCs. All replicas of a content can be lost in a large-scale disaster. Critical data must be quickly evacuated from DCs in the disaster region to safe DCs prior to the disaster. Rapid data evacuation. Receive warning of an oncoming disaster from various sensors and monitors in their network and/or from government or intelligence agencies. Depending on the type of disaster (e.g., earthquake, hurricane or weapons of mass destruction (WMD)), do the following prediction. Disaster zone Evacuation deadline Potential damage in the infrastructure Before deadline, quickly evacuate as much critical data as possible. 4

5 Background (contd.) Safe data transfers Possible independent disasters may compromise the process of data evacuation. Risk of node/link failures must be considered to ensure safe data transfers. How to objective a tradeoff between evacuation time and risk of data loss during evacuation. 5

6 Risk-Aware Rapid Data Evacuation For Large-Scale Disasters (contd.) Time delay Path-computation delay. Connection-setup delay. Data-transmission delay. Data-propagation delay. Notations Distance of path: l. Number of hops on path: n. Bandwidth of path: B p Propagation delay per unit distance: μ. Processing delay: η. Switch configuration delay: β. Assume the same propagation delay for data and control messages. 6

7 Risk-Aware Rapid Data Evacuation For Large-Scale Disasters (contd.) Equations Connection-setup delay Control-message processing delay: (n +1) η Control-message propagation delay: l μ Switch-configuration delay: (n + 1) β Transmission delay: F c B p Propagation delay: l μ 7

8 Risk-Aware Rapid Data Evacuation For Large-Scale Disasters (contd.) The risk of node/link failures rrrr l m /rrrr n m are the probabilities of link l or node n being damaged due to disasterm M. l L,m M(1 rrrr m l ) n N,m M (1 rrrr m n ) computes the probability that all links and nodes are not damaged by any disaster. Path failure probability: rrrr p = 1 l L,m M(1 rrrr m l ) n N,m M (1 rrrr m n ). 8

9 Risk-Aware Rapid Data Evacuation For Large-Scale Disasters (contd.) Example Transfer content c1 from node C to safe DCs A or D as fast as possible. Failure probabilities of links C-D and A-C is assumed as 0.6 and 0.2. Select destination DC D Evacuation time is less but the risk along path C-D is higher. Select destination DC A Evacuation time is high but the risk along path C-D is less. 9

10 Risk-Aware Rapid Data Evacuation For Large-Scale Disasters (contd.) Problem statement Objective: Achieve an optimal tradeoff between evacuation time and risk of data loss during evacuation. Given inputs: N is the set of nodes and L is the set of links. Physical topology G = (N, L). Set of possible disaster zones M. Set of predicted WMD attack zones WW. Set of DCs D. Storage capacity S d. Set of locally hosted contents C d. Number of replicas of content c R c. Importance metric of content c α c. 10

11 Risk-Aware Rapid Data Evacuation For Large-Scale Disasters (contd.) Problem statement Input: Size of content c F c. Residual link capacity B l, l L. Evacuation deadline T. Set of k-shortest paths R ii for each node pair (i, j), i, j N. Constraint: Available link capacity is limited. Data-transfer delay to be upper bounded by evacuation deadline. Different paths can be used in parallel for different connections if paths do not overlap. 11

12 Heuristic Algorithm Disaster mapping: Get a set of DCs D ww D, ww WW. // Datacenters in predicted disaster zone Get a set of DCs D ooo_ww D, D ooo_ww = D D ww. // Datacenters outside predicted disaster zone Content selection: For each DC d D ww Get a set of contents, C d, d D ww, C ww = C ww C d, ww WW. // Contents in predicted disaster zone For each content c C ww If all replicas of content c are in the disaster zone WW then Put c in a content list C EEE and get set of DCs D c, which host the replicas of c. // C EEE is a set of contents to be evacuated Sort C EEE based on α c in a descending order. 12

13 Heuristic Algorithm Destination DC selection and path delay and risk computation: For each content c C EEE For each d D ooo_ww, if F c S d then put d into list of available DCs D AAA // D AAA is a set of safe DCs with available storage Get set of k-shortest paths R ii, i D c,j D AAA For each path p R ii Obtain total delay ddddd p and risk rrrr p to calculate general cost CCCC p CCCC p = ddddd p + φ rrrr p Set path p which has minimum cost CCCC p as final solution 13

14 Performance Evaluation Simulation conditions Network topology: 24-node USNET topology. Disaster type: WMD attacks. A predicted WMD zone: WM A set of 10 possible independent disaster zones: M. Blue nodes represent safe DCs. Red nodes 6, 9, and 12 are the DCs in the WMD zone WM. 14

15 Performance Evaluation (contd.) Large storage capacity ranging from 10TB to 100TB (average occupation is assumed to be 40%). Residual link capacity is assumed to range from 500Gbps to 1Tbps (network is assumed to have 30% utilization). Number of contents is assumed to be 300. Size of each content is randomly generated within the range of [100GB, 200GB]. The contents are uniformly distributed among different DCs with the number of replicas ranging from 2 to 4. Contents are randomly assigned the importance metric α c on a scale from 1 to 10. Processing delay, propagation delay, and switch configuration delay to be 10 μs, 5 μs/km, 15 ms. 15

16 Performance Evaluation (contd.) Total risk of data loss and evacuation time Time consumption (s) Weight factor ϕ Shows the time consumption and the total risk of data loss for the RA-RDE algorithm with an increasing weight factor φ. We see that RA-RDE can significantly reduce the total risk with an increasing φ. This is reasonable since a larger weight factor can lead to higher risk reduction in the general cost Total risk of data loss 16

17 Performance Evaluation (contd.) Performance comparison Result gap compared with LD-RDE 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% Time consumption Total risk of data loss 0.00% Weight factor ϕ Compare the performance of RA-RDE with the LD-RDE algorithm which selects the least-delay paths for data evacuation without considering risk. It should be noted that the RA-RDE algorithm is equivalent to the LD-RDE scheme when φ = 0. We see that, when φ = 3, our approach reduces the total risk by 20% and needs less than 10% additional time consumption. 17

18 Performance Evaluation (contd.) Time consumption with an increasing number of contents Time consumption (s) Number of contents 18 NDE RA-RDE LD-RDE Compare time consumption of the two rapid evacuation approaches with the nearest data evacuation (NDE) approach which evacuates data only to the nearest DC. With an increasing number of contents from 100 to 500 and φ = 3, we can see that RA-RDE performs close to LD-RDE and is much better than NDE, which verifies its time efficiency.

19 Conclusion To balance performance between time consumption and total risk of data loss, we defined a general cost considering path delay and path risk using a weight factor. We develop a risk-aware rapid data evacuation scheme for largescale disasters in optical cloud networks. Results show that proposed approach significantly reduces total risk with minimal addition time consumption. Time consumption close to LD-RDE under different number of contents. 19

20 Future Work Try to propose an MILP model to solve the rapid data evacuation problem by using time slot Time slot contiguity Time slot continuity C1 C2 12 Time slot T Investigate an new heuristic algorithms with the ICC paper of Wu Yu 20

21 Thank you for your attention! Presenter: Yongcheng (Jeremy) Li PhD student, School of Electronic and Information Engineering, Soochow University, China Group Meeting, Friday, August 26,